Multi-Sample Aligned Change-Point Detection Using Penalized Test Statistics

• Activity Number: 161
• Type: Topic Contributed
• Date/Time: Monday, August 10, 2015 : 10:30 AM to 12:20 PM
• Sponsor: IMS
• Author(s): Hock Peng Chan* and Guenther Walther
• Companies: National University of Singapore and Stanford University
• Keywords: Berk-Jones ; higher criticsm ; scan statistics ; sparse mixtures

Abstract

There have been a lot of recent interest in the detection of aligned change-points in parallel streams of data. Applications include communications, disease surveillance, engineering and hospital management. These are interesting statistical problems as they bring in an additional multi-sample dimension to traditional change-point detection works that consider only a single stream of data. We show here the critical boundary separating detectable from non-detectable change-points, and construct penalized Berk-Jones and higher-criticism test statistics that achieve optimal detectability. Extensions to template-matching problems will also be discussed.